, k = 1, . . . , n−1
and apply each time Lemma 3.1. After at most n−1 steps we get a matrix
C = (cij)∈M(n), detC = detA cjn=cnj = 0, j = 1, . . . , n−1, cnn 6= 0
Z
Rn
f(Cx)dx= Z
Rn
f(Ax)dx
This procedure we apply step by step to all further columns and rows. After at most n2 steps we find the desired diagonal matrixD with (3.10). By (1.11)
Z
Rn
f(y)dy=|detnD|
Z
Rn
f(Dx)dx=
=|detnA|
Z
Rn
f(Ax)dx.
Combining last identity with (3.9) yields (3.8).
4 Some properties of bi-Lipschitz mappings
Throughout this chapter, let G ⊂ Rn be an open set and, if not otherwise stated, letu:G→Rn be abi-Lipschitz mapping. Let
N ⊂G, |N|= 0
such thatu is differentiable onG\N (compare Theorem 1.5).
Lemma 4.1 Let x0 ∈ G \ N. Then there is δ = δ(x0, G, N, u) > 0 such that Bδ(u(x0))⊂u(G).
Proof.
a) By Theorem 1.5 there is ε0 >0 such that Bε0(x0)⊂G and there is ϕ:Bε0(0)→Rn, lim
Bε0(0)3ξ→0 ξ6=0
ϕ(ξ) kξk = 0 such that
(4.1) u(x) = u(x0) +u0(x0)(x−x0) +ϕ(x−x0) ∀x∈Bε(x0).
We choose 0< ε < ε0 such that
(4.2) kϕ(ξ)k ≤ L1
4 kξk ∀ξ∈Bε(0)
(where 0 < L1 ≤ L2 by Theorem 1.6). We set δ := L41ε > 0 and for y ∈ Bδ(u(x0))
(4.3)
u(x, t) :=tu(x) + (1−t) [u(x0) +u0(x0)(x−x0)]−y for (x, t)∈Bε(x0)×I.
It is immediately seen that hypotheses (i) and (ii) of Theorem 2.5 hold (with G replaced by Bε(x0)). Forkx−x0k=ε, t∈[0,1] by (4.1)
u(x, t) = u0(x0)(x−x0) +tϕ(x−x0) +u(x0)−y By (1.29) and (4.2)
(4.4)
ku(x, t)k ≥ ku0(x0)(x−x0)k − kϕ(x−x0)k − ku(x0)−yk
≥L1·ε− L41ε− L41ε= L21ε = 2δ >0
∀kx−x0k=ε, ∀t∈I Let U :={y∈Rn :kyk> δ}. Thenu(x, t)∈U for (x, t)∈∂Bε(x0)×I.
By Theorem 2.5 (applied to Bε(x0)) for f ∈Cc0(Rn) with f¯
¯U = 0 Z
Bε(x0)
f(u(x)−y) detnu0(x)dx= (4.5)
= Z
Bε(x0)
f(u(x0) +u0(x0)(x−x0)−y)·detnu0(x0)dx
b) Assume now that there is no x∈Bε(x0) such that u(x) =y. Then min
n
ku(x)−yk¯
¯x∈Bε(x0) o
=:σ > 0.
For x=x0 we see because of y∈Bδ(u(x0))
(4.6) σ ≤ ku(x0)−yk< δ.
Let
(4.7) ϕ(t) :=
(¡σ
2 −t¢
for 0≤t ≤ σ2 0 for t≥ σ2
and f(z) := ϕ(kzk) for z ∈ Rn. f ∈ Cc0(Rn) and because of (4.6) it vanishes on U. Further, f(u(x)−y) = 0 for x ∈ Bε(x0) and therefore the integral at the left of (4.5) is zero.
On the other hand, detnu0(x0) 6= 0, and there exists a unique z ∈ Rn such that y−u(x0) = u0(x0)·z. By (1.29)
L1ε
4 =δ >ky−u(x0)k=ku0(x0)·zk ≥L1kzk, hence kzk< 4ε and
x1 :=x0+z ∈Bε/4(x0),
f(u(x0) +u0(x0)(x1 −x0)−y) =f(0) = σ 2. By continuity there is τ > 0 such that Bτ(x1)⊂Bε(x0) and
f(u(x0) +u0(x0)(x−x0)−y)≥ σ
4 ∀x∈Bτ(x1).
Then
|detnu0(x0)|
Z
Bε(x0)
f(u(x0) +u0(x0)(x−x0)−y)dx≥
≥ σ
4|Bτ(x1)| |detnu0(x0)|>0, a contradiction.
Corollary 4.2 Let N∗ := u(N) ⊂ u(G), |N∗| = 0 (Lemma 1.7). Then there are open sets V, V∗ ⊂Rn such that
u(V) =V∗
G\N ⊂V ⊂G, u(G)\N∗ ⊂V∗ ⊂u(G).
Proof. For x∈G\N let Bδx(x) be according Lemma 4.1 and set V∗ := [
x∈G\N
Bδx(x)
Then V∗ ⊂ Rn is open, V∗ ⊂ u(G) and V :=u−1(V∗)⊂ G is open too. Clearly by construction,G\N ⊂V and because u is injective, u(G)\N∗ ⊂V∗.
Lemma 4.3 For each x0 ∈G\N there is δ =δ(x0, G, N, u)>0 such that i) Bδ(u(x0))⊂u(G)
ii) sgn detnu0(x) = sgn detnu0(x0) for all x∈u−1(Bδ(u(x0))\N) iii) for every f ∈Cc0(Rn) with suppf ⊂Bδ(u(x0))
(4.8)
Z
u(G)
f(y)dy= Z
G
f(u(x))|detnu0(x)|dx
Proof.
a) We proceed similarly as in the proof of Lemma 4.1. We choose ε > 0 and δ = L41ε as there. Now we consider
u(x, t) := tu(x) + (1−t) [u(x0) +u0(x0)(x−x0)] ∀(x, t)∈Bε(x0)×I.
Again hypotheses i) and ii) of Theorem 2.5 hold with G replaced by Bε(x0).
By (4.1)
u(x, t) = u(x0) +u0(x0)(x−x0) +tϕ(x−x0).
For kx−x0k=ε,t ∈I, by (1.29) and (4.2)
(4.9) ku(x, t)−u(x0)k ≥L1kx−x0k − L1
4 kx−x0k= 3
4L1·ε= 3δ Let now
U0 :=©
y ∈Rn¯
¯ky−u(x0)k> δª By (4.9)
u(x, t)∈U0 for (x, t)∈∂Bε(x0)×I
and by Theorem 2.5 for all f ∈Cc0(Rn) with suppf ∈Bδ(u(x0))
(4.10)Z for (4.11). Suppose in addition, that h≥ 0, then f ≥0 too. We consider the case σ= 1. Then by (4.11)
Since |detu0(x)| ≤C ∀x∈Ω, Z
Ω
|χM(x)− |hk(x)|| |detnu0(x)|dx≤
≤C Z
Ω
|χM(x)− |hk(x)||dx.
By (4.12) Z
Ω\N
χM(x) detnu0(x)dx= lim
k→∞
Z
Ω\N
|hk(x)|detnu0(x)dx≥0
On the other hand χM(x) detnu0(x) ≤ 0, for x ∈ Ω\N. Therefore |M| = 0 and detnu0(x)>0 a. e. in Ω\N, σ·detnu0(x) =|detnu0(x)|and (4.7) follows from (4.11). The case σ=−1 is treated analogously.
Lemma 4.4 Let G ⊂ Rn be in addition connected. Then either detnu0(x) > 0 or detnu0(x)<0 ∀x∈G\N.
Proof. Let xi ∈ G \ N(i = 0,1) x0 6= x1. Then there is a continuous curve γ :I →G such that γ(i) =xi(i= 0,1). Let
¯ γ :=©
γ(t)¯¯t∈[0,1]ª .
Then ¯γ ⊂ G is compact, hence d := dist(¯γ, ∂G) >0 (if ∂G 6= φ; set d := 1 if G = Rn). We choose δi > 0 according Lemma 4.3 such that Bδi(xi) ⊂ u(G) (i = 0,1).
Let
δ:= 1 2min
µ
δ1, δ2, d,kx0−x1k 2
¶
>0 Then
G0 :=©
x∈G¯
¯dist(x,γ)¯ < δª is a bounded open set subset ofG, G0 ⊂⊂G. We set
u(x, t) := u(x)−u(γ(t)) for (x, t)∈G¯0×I.
Clearly, ∂G0 ⊂ ©
x∈G¯
¯dist(x,γ) =¯ δª
. By compactness of ¯γ for every x ∈ ∂G0 there exists t0 ∈I such that
kx−γ(t0)k=δ= inf©
kx−γ(t)k¯
¯t∈Iª . Then fort ∈I,x∈∂G0
(4.13) ku(x, t)k=ku(x)−u(γ(t))k ≥L1kx−γ(t)k ≥L1kx−γ(t0)k=L1·δ
Let
U :=
½
y∈Rn¯¯dist(y, u(¯γ))> L1δ 2
¾
By (4.13) u(∂G0×I)⊂U. Let nowϕ∈Cc0(Rn), suppϕ⊂Bδ(0) and Z
Bδ(0)
ϕ(y)dy = 1.
By Theorem 2.5 (applied toG0) Z
G0
ϕ(u(x,0)) detnu0(x,0)dx = Z
G0
ϕ(u(x,1)) detnu0(x,1)dx that is
Z
G0
ϕ(u(x)−u(x0)) detnu0(x)dx = Z
G0
ϕ(u(x)−u(x1)) detnu0(x)dx
Sinceϕ(u(x)−u(xi)) vanishes forx /∈u−1(Bδ(u(xi)), the domain of integration may be replaced by these sets respectively. According Lemma 4.3 ii)
sgn detnu0(x) = sgn detnu0(xi) =:σi ∀x∈u−1(Bδu(xi))\N, i= 1,2 Therefore
σ0 Z
u−1(Bδ(u(xo)))
ϕ(u(x)−u(x0))|detnu0(x)|dx=
=σ1 Z
u−1(Bδ(u(x1)))
ϕ(u(x)−u(x1))|detnu0(x)|dx
We apply Lemma 4.3 to each of this integrals and to the mapsvi(x) :=u(x)−u(xi) and we get
σ0 Z
Bδ(0)
ϕ(y)dy =σ1 Z
Bδ(0)
ϕ(y)dy
and thereforeσ0 =σ1. Since xi ∈G\N (i= 0,1) had been arbitrary (x0 6=x1), the claim is proved.
The next result is a special case of a famous theorem by L. E. J. Brouwer. This theorem guarantees that a continuous, locally injective mappingf :G→Rn, where G⊂ Rn is open, is an open mapping, i. g. it maps open subsets U ⊂ G onto open subsets f(U)⊂Rn. Especially, f(G) is open. Here, locally injective means that to every x ∈ G there exists a neighborhood Ux ⊂ G such that f¯
¯Ux : Ux → f(Ux) is injective (see e.g. [Dei], Theorem 4.3).
Clearly, locally bi-Lipschitz mappings are continuous and locally injective.
Theorem 4.5 Let G ⊂ Rn be open and let u : G → Rn be a locally bi-Lipschitz mapping. Then u is open.
Proof.
a) We prove first that for each x ∈G there is εx >0 such that Bεx(x) ⊂G and there is δx > 0 such that Bδx(u(x)) ⊂ u(Bεx(x)). Without loss of generality let x0 = 0∈G and u(0) = 0. Otherwise consider
˜
u(x) := u(x−x0)−u(x0) for x ∈ G˜ := ©
y+x0¯¯y∈Gª
. Then there is ε > 0 such that Bε(0) ⊂ G and u¯
¯Bε(0) is bi-Lipschitz (compare Definition 1.3, part 4). Clearly, by continuity, there is a unique bi-Lipschitz extension to Bε(0) : 0< L1 ≤L2
(4.14) L1kx−x0k ≤ ku(x)−u(x0)k ≤L2kx−x0k ∀x, x0 ∈Bε(0) Consider y∈Bδ(0), where δ:= L21ε >0 and consider
u(x, t) := u(x)−ty for (x, t)∈Bε(0)×I.
Then for (x, t)∈∂Bε(0)×I by (4.14) and u(0) = 0
(4.15) ku(x, t)k ≥ ku(x)k − kyk ≥L1kxk − kyk> L1·ε− L1ε
2 = L1ε 2 =δ Suppose that y /∈u
³ Bε(0)
´
. Then σ:= min
n
ku(x)−yk¯
¯x∈Bε(0) o
>0.
Clearly,
σ ≤ ku(0)−yk=kyk< δ = L1ε 2 .
Consider again ϕ defined by (4.7) and f(z) :=ϕ(kzk) forz ∈Rn. Let U :=©
y∈Rn¯
¯kyk> δª . By (4.15) u(∂Bε(0)×I)⊂ U, f ∈ Cc0(Rn) and f¯¯
U = 0. By Theorem 2.5 we see
(4.16)
Z
Bε(0)
ϕ(ku(x)−yk) detu0(x) = Z
Bε(0)
ϕ(ku(x)k) detu0(x)dx
Then ϕ(ku(x)−yk) = 0 ∀x ∈ Bε(0) and the left integral vanishes. On the other hand u(0) = 0, therefore ϕ(ku(0)k) = σ2. Then there is 0< ε0 < ε such that
ϕ(ku(x)k)≥ σ
4 >0 ∀x∈Bε0(0).
Since Bε(0) is a domain, by Lemma 4.4 detnu0(x) is of constant sign on Bε(0)\N. Without loss of generality let sgn detnu0(x) = 1 forx∈Bε(0)\N. By (4.16)
(4.17) 0 =
Z
Bε(0)
ϕ(ku(x)k) detnu0(x)dx≥ σ 4
Z
Bε0(0)
detnu0(x)dx.
Because of the first inequality in (1.19), detnu0(x) 6= 0 ∀x ∈ Bε0(0)\N. If we suppose R
Bε0(0)
detnu0(x)dx = 0, we would conclude (observe detnu0(x) ≥ 0 inR Bε0(0)) detnu0(x) = 0 a. e. in Bε0(0), a contradiction. Therefore
Bε0(0)
detnu0(x)dx > 0 and we get by (4.17) a contradiction. Since y ∈ Bδ(0) was arbitrary we see Bδ(0) ⊂Bδ(0)⊂u(Bε(x)).
b) By part a) of proof, for everyx∈Gthere isδx>0 such thatBδx(u(x))⊂u(G).
Then S
x∈G
Bδx(u(x)) is open in Rn and clearly u(G)⊂ [
x∈G
Bδx(u(x))⊂u(G).
Therefore u(G) is open in Rn.
c) If G0 ⊂ G is any open set, then we apply parts a) and b) of proof to G0 in place of G and we see by b) that u(G0) is open in Rn.
Theorem 4.6 Let G ⊂ Rn be open and let u : G → Rn be a locally bi-Lipschitz mapping. Then
i) G∗ :=u(G) is open
ii) For every x0 ∈G there are open neighborhoods Ux0 ⊂ G and Vy0 ⊂G∗(y0 :=
u(x0)) such that u¯
¯Ux0 : Ux0 → Vy0 is bijective and there is a set M ⊂ Vy0,
|M|= 0, such that
v :Vy0 →Ux0, v :=
³ u¯
¯Ux0
´−1
is differentiable on Vy0 \M.
Proof. Due to Theorem 4.5, G∗ is open. There is a neighborhood Ux0 such that u¯
¯Ux0 : Ux0 → Rn is a bi-Lipschitz mapping, hence u : Ux0 → Vy0 := u(Ux0) is bijective andVy0 is open by Theorem 4.5. Due to Rademacher’s theorem (Theorem 1.4) there isM ⊂Vy0, |M|= 0 such that v :Vy0 \M →Uy0 is differentiable.
Theorem 4.6 is a generalization of the classical theorem on local diffeomorphisms.
Theorem 4.7 Let G ⊂ Rn be open and let u : G → Rn be an injective, locally bi-Lipschitz mapping. Then
i) G∗ :=u(G) is open,
u:G→G∗ is bijective and differentiable on G\N, |N|= 0.
ii) There is a setM ⊂G∗, |M|= 0 such thatv :=u−1 is differentiable onG∗\M. Proof. By Theorem 4.5, G∗ is open and because of global injectivity, u :G→G∗ is bijective. Let (G∗k)⊂G∗ be an exhausting sequence for G∗:
G∗k ⊂⊂G∗k+1 ⊂⊂G∗ ∀k ∈N, [∞
k=1
G∗k =G∗
Due to Theorem 4.6 for each y0 ∈G∗k there is an open Vy0 and a set My0, |My0|= 0 such that v¯
¯Vy0\M is differentiable. By compactness of G∗k there are y(k)i ∈ G∗k (i = 1, . . . , mk) such that G∗k ⊂ mSk
i=1
Vy(k)
i , Mi(k) ⊂ Vy(k)i , |Mi(k)| = 0 and v¯
¯Vyi(k)\Mi(k) is differentiable. Then v¯
¯G∗k\Mk is differentiable, whereMk := mSk
i=1
Mi(k), |Mk| = 0. Let M := S∞
k=1
Mk ⊂G∗,|M|= 0. Then v is differentiable onG∗\M. As a further corollary we derive
Theorem 4.8 (implicit function theorem) Letm, n∈Nand letUm ⊂Rm, Un⊂Rn be open sets, U :=Um×Un ⊂Rn+m. Let
f :U →Rn and suppose that
i) there is a∈Um, b∈Un such that
f(a, b) = 0 ii) there exists L >0 such that
kf(x, y)−f(x0, y0)kn ≤L(kx−x0km+ky−y0kn) ∀x, x0 ∈Um, ∀y, y0 ∈Un
iii) there exists K >0 such that
kf(x, y)−f(x, y0)kn ≥Kky−y0kn ∀y, y0 ∈Un ∀x∈Um. Then there exists an open neighborhoodVm ⊂Um of a and a Lipschitz mapping
g :Vm →Rn such that
10 (x, g(x))∈U ∀x∈Vm
20 there is a subset Nm ⊂Vm, |Nm|= 0 and g is differentiable on Vm\Nm 30 g(a) = b and f(x, g(x)) = 0 ∀x∈Vm
40 ©
(x, y)∈Vm×Un¯
¯f(x, y) = 0ª
=©
(x, g(x))¯
¯x∈Vmª Proof.
a) Let h:U →Rm+n be defined by
hi(x, y) :=αxi for i= 1, . . . , m
(x, y)∈Um×Un =U hm+k(x, y) :=fk(x, y) for k= 1, . . . , n
where α=£1
2K2+L2¤1
2 >0.
For (x, y), (x0, y0)∈U
kf(x, y)−f(x0, y0)kn≥¯¯kf(x, y)−f(x, y0)kn− kf(x, y0)−f(x0, y0)kn¯¯. Then by ii) and iii) (for ε >0: 2|a·b| ≤εa2+ε−1b2; chooseε:= 12)
kf(x, y)−f(x0, y0)k2n ≥(kf(x, y)−f(x, y0)kn− kf(x, y0)−f(x, y0)kn)2
≥ 1
2kf(x, y)−f(x, y0)k2n− kf(x, y0)−f(x, y0)k2n
≥ 1
2K2ky−y0k2n−L2kx−x0k2m Hence
kh(x, y)−h(x0, y0)k2n+m=α2kx−x0k2m+kf(x, y)−f(x0, y0)k2n
≥(α2−L2)kx−x0k2m+ 1
2K2ky−y0k2n
≥ 1
2K2k(x, y)−(x0, y0)k2n+m
and therefore
kh(x, y)−h(x0, y0)kn+m ≥L1k(x, y)−(x0, y0)kn+m, ∀(x, y),(x0, y0)∈U (4.18)
L1 := 1 2K ·√
2>0.
kh(x, y)−h(x0, y0)kn+m ≤αkx−x0km+L(kx−x0km+ky−y0km) (4.19)
≤L2k(x, y)−(x0, y0)kn+m ∀(x, y),(x0, y0)∈U where L2 :=α+ 2L >0.
Because of (4.18), (4.19)h:U →Rnis a bi-Lipschitz mapping and by Theorem 4.7 U∗ :=h(U) is open. Let
v :=h−1 :U∗ →U and by (4.18), (4.19) for z ∈Rn, ω ∈Rm, (z, ω)∈U∗
L−12 k(z, ω)−(z0, ω0)kn+m ≤ kv(z, ω)−v(z0, ω0)kn+m
(4.20)
≤L−11 k(z, ω)−(z0, ω0)kn+m ∀(z, ω),(z0, ω0)∈U∗. b) Since h(a, b) = (αa,0)∈U∗ and U∗ is open, there isδ >0 such that
Bδ(αa,0) =©
(z, ω)∈Rn+m¯
¯k(z, ω)−(αa,0)kn+m < δª
⊂U∗. Let
Bm :=
½
z ∈Rm¯
¯kz−αakm < δ
√2
¾
Bn:=
½
ω ∈Rn¯
¯kωkn< δ 2
¾
Obviously (αa,0)∈Bm×Bn ⊂U∗. Let Vm :=
½
x∈Rm¯
¯kx−akm< δ α√
2
¾
Observe that x∈Vm if and only if αx∈Bm. We write v =¡
˜ v,v˜˜¢
, where
˜
v := (v1, . . . , vm), v˜˜:= (vm+1, . . . , vm+n).
(z, ω)∈U∗ if and only if there is a unique (x, y)∈U such that (z, ω) = h(x, y) = (αx, f(x, y)).
Therefore z =αx,ω =f(x, y), hence
˜
v(z, ω) =x, v(z, ω) =˜˜ y Since h(a, b) = (αa, f(a, b)) = (αa,0),
˜˜
v(αa,0) =b.
If (x, ω)∈Vm×Bn, then (αx, ω)∈Bm×Bn and
(αx, ω) = h(v(αx, ω)) = (αx, f(x,˜˜v(αx, ω)), hence
ω =f(x,v˜˜(αx, ω)) If ω = 0, 0∈Bn, x∈Vm, then
¡x,˜˜v(αx,0)¢
∈U.
We set
(4.21) g(x) := ˜˜v(αx,0) for x∈Vm g is continuous, (x, g(x))∈U and
0 =f(x, g(x)) ∀x∈Vm.
Further, g(a) = ˜˜v(αa,0) = b. If (x, y) ∈ Vm ×Un ⊂ U and f(x, y) = 0, then h(x, y) = (αx,0) ∈ U∗, therefore v(αx,0) = (x,v(αx,˜˜ 0)) = (x, g(x)). Then trivially 40 is true. By (4.20) for x, x0 ∈Vm
kg(x)−g(x0)kn =k˜˜v(αx,0)−˜˜v(αx0,0)kn≤
≤ kv(αx,0)−v(αx0,0)kn+m ≤L−11 kα(x−x0)km =
=L−11 αkx−x0km
By Rademacher’s theorem there is Nm ⊂ Vm, |Nm|m = 0, such that g is differentiable on Vm\Nm.
Corollary 4.9 Let the hypotheses of Theorem 4.8 hold true. Then there is N ⊂ Rn+m, |N|n+m = 0 such that f is differentiable on U \N. For (x, y)∈U\N let
(Dyf)(x, y) =
∂f1
∂y1(x, y) · · · ∂y∂fn1(x, y)
... ...
∂fn
∂y1(x, y) · · · ∂f∂yn
n(x, y)
Then
(4.22) Kkηkn ≤ k(Dyf)(x, y)ηkn≤Lkηkn ∀η∈Rn, ∀(x, y)∈U \N and
rank (Dyf)(x, y) = n ∀(x, y)∈U\N.
Proof. The proof is completely analogous to the proof of Theorem 1.5
Clearly f and g are a.e. differentiable and f(x, g(x)) = 0 ∀x ∈ Vm too. At points x ∈ Vm\Nm such that (x, g(x))∈/ N ⊂ Rn+m (N denoting the set where f is not differentiable), then the chain rule may be applied tof(x, g(x)). But it may happen that (x, g(x))∈N ∀x∈Vm. A simple example:
LetU :=©
(x, y)∈R2¯
¯|x|< 12,|y|<1ª
f(x, y) :=
(1 +x)y for y≥0
(x, y)∈U (1 +x)2y for y <0
Thenf(0,0) = 0,
kf(x, y)−f(x0, y0)k1 ≤3(kx−x0k1+ky−y0k1) kf(x, y)−f(x, y0)k1 ≥ 1
2ky−y0k1 ∀|x|< 1
2 ∀|y|<1.
All hypotheses of Theorem 4.8 hold true. Let Vm :=©
x∈R¯
¯|x|< 12ª
, g : Vm →R, g(x) := 0. Then f(x, g(x)) = 0 ∀x∈Vm. The set where f is not differentiable is
N :=©
(x, y)∈U¯
¯y= 0ª
⊂R2, |N|2 = 0 and ©
(x, g(x))¯
¯x∈Vmª
=N. The chain rule is not applicable.